[BE-0281] feat(e2e): add real on-device iOS actuation coverage by 0x0c · Pull Request #1181 · bajutsu-e2e/bajutsu

0x0c · 2026-07-17T13:49:52Z

Summary

Wires real on-device actuation into the iOS E2E lane (ios-e2e.yml) beyond the conformance job's tap, closing the CI-wiring gap BE-0281 identifies: the showcase already carries gestures.yaml, device.yaml, and push.yaml, but they ran only in local-only Makefile targets, never CI. This PR moves BE-0281 to Implemented, with scope narrowed from the original design (see Scope below).

What changed

.github/workflows/ios-e2e.yml — adds one new non-gating job, actuation (idb), driving navigation.yaml (tap + back) and device.yaml + push.yaml (setLocation, clipboard, push) over idb on the Stable launch tab — the only surface idb can reach. Also wires two more runs into the existing xcuitest (multi-touch) job: search.yaml for the runner channel's /type endpoint and notices.yaml for /swipe + /back, reusing that job's already-built runner rather than a new job. actuation stays out of the E2E gate's needs:, like golden/visual, landing as signal first per the Simulator lane's flakiness history (BE-0218).
docs/ci.md / docs/ja/ci.md — update the iOS lane's job inventory (six jobs → seven) to describe the new actuation (idb) job and the xcuitest (multi-touch) job's expanded scope.
roadmaps/BE-0281-ios-on-device-actuation-coverage/*.md — flip Status to Implemented, check off the Progress items, and log what shipped and why the scope narrowed.
roadmaps/README.md / roadmaps/README-ja.md — regenerated via make roadmap-index (BE-0281 moves from Proposal to Implemented).

Scope

Narrowed from the original proposal. idb cannot tap the native tab bar at all — SwiftUI's TabView collapses into one opaque group in idb's describe-all, with no per-tab children (SPEC.md §3, BE-0107) — confirmed on-device before implementation (a tap: {label: "Log"} step returned 一致なし/no-match). The proposal's idb interaction job assumed gestures.yaml and similar tab-crossing scenarios would run over idb; they cannot, since every one of them opens with a tab tap. The only existing scenario idb can drive end-to-end is navigation.yaml (Stable launch tab only), covering tap + back.

Consequently:

idb's type / swipe / scroll / longPress / doubleTap stay proven on-device only via the XCUITest runner channel (this PR's search.yaml / notices.yaml wiring) or mocked unit tests — not via the idb companion itself. Closing that specific gap needs a new tab-free showcase screen (idb has no surface today with a text field, scrollable list, or gesture targets outside the tab bar). Left for a follow-up roadmap item.
Device control (setLocation, clipboard, push) is fully covered as proposed — it lives entirely on the Stable launch tab, which idb reaches.
XCUITest actuation (/type, /swipe, /back) is fully covered as proposed, reusing the existing shared search.yaml/notices.yaml scenarios rather than authoring a new file.

Prime-directive compliance

AI never judges — no LLM is consulted; every new job is a deterministic bajutsu run against machine-checkable expect assertions.
Determinism first — no fixed sleep; the new job reuses the same BAJUTSU_MIN_WAIT_TIMEOUT condition-wait floor pattern as smoke/golden/visual.
App-agnostic — no per-app logic added; the new job runs existing showcase scenarios through the existing bajutsu-e2e composite action.

Verification

make check
  format-check / lint / lint-docstrings / lint-imports / lint-sh / lint-actions / lint-js /
  lint-roadmap / lock-check / typecheck: all passed
  4369 passed, 3 skipped, coverage 92.37% (floor 89%)

actionlint and YAML-parse checks pass on the modified workflow. The XCUITest half of the new coverage (search.yaml for /type, notices.yaml for /swipe + /back) was additionally verified end-to-end on a local Simulator + Xcode 26.6 toolchain before landing — both scenarios PASS over --backend xcuitest. The idb half (navigation.yaml, device.yaml, push.yaml) could not be locally verified the same way (the local idb_companion build is too old to read the iOS 26 accessibility tree) — the job's actual on-device behavior will be observed on this PR's CI run, which is exactly why it starts non-gating.

🤖 Generated with Claude Code

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

…ntrol on-device - ios-e2e.yml: notices.yaml only uses `/swipe` + `/back` steps (no `/scroll` endpoint exists on the runner channel), so drop the stray "+ `/scroll`" from the job-header comment describing it. - docs/architecture.md and its docs/ja/architecture.md mirror: credit idb's `back` and device-control (setLocation/clipboard/push) actuation as now confirmed on-device via the `actuation (idb)` job (BE-0281), per CLAUDE.md's BE-0113 rule to keep architecture.md in step with behavior in the same PR. Addresses review feedback on PR #1181. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

Address round-3 review feedback on PR #1181: the actuation (idb) job's three bajutsu-e2e steps (navigation/device/push) are independent scenarios bundled into one job for cost reasons, but a plain step failure skipped every step after it — hiding the other scenarios' pass/ fail signal and artifacts. Mark each step continue-on-error and add a final step that re-fails the job if any of the three failed, so all three always get a chance to run while the job-level result stays accurate. Also fixes documentation drift flagged by the same review round: the roadmap's References list still cited gestures.yaml/controls.yaml (never actually wired over idb) instead of the scenarios this PR wired (navigation/device/push/search/notices.yaml), and architecture.md overclaimed the actuation (idb) job "confirmed" back/device-control on-device before a passing run had been observed. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

…ture.md antecedent Address two open review comments on PR #1181: - Bump ios-e2e.yml xcuitest-gestures timeout-minutes 25 -> 30 to restore headroom after BE-0281's ~13min -> ~21min runtime growth on a gating job. - Fix 'runs it' -> 'runs both' in docs/architecture.md's idb entry: the bullet describes two things (back and device control), so the singular pronoun misread. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NgEYipfg6U2xLeCAxTTNF7

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

…ntrol on-device - ios-e2e.yml: notices.yaml only uses `/swipe` + `/back` steps (no `/scroll` endpoint exists on the runner channel), so drop the stray "+ `/scroll`" from the job-header comment describing it. - docs/architecture.md and its docs/ja/architecture.md mirror: credit idb's `back` and device-control (setLocation/clipboard/push) actuation as now confirmed on-device via the `actuation (idb)` job (BE-0281), per CLAUDE.md's BE-0113 rule to keep architecture.md in step with behavior in the same PR. Addresses review feedback on PR #1181. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

Address round-3 review feedback on PR #1181: the actuation (idb) job's three bajutsu-e2e steps (navigation/device/push) are independent scenarios bundled into one job for cost reasons, but a plain step failure skipped every step after it — hiding the other scenarios' pass/ fail signal and artifacts. Mark each step continue-on-error and add a final step that re-fails the job if any of the three failed, so all three always get a chance to run while the job-level result stays accurate. Also fixes documentation drift flagged by the same review round: the roadmap's References list still cited gestures.yaml/controls.yaml (never actually wired over idb) instead of the scenarios this PR wired (navigation/device/push/search/notices.yaml), and architecture.md overclaimed the actuation (idb) job "confirmed" back/device-control on-device before a passing run had been observed. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

…ture.md antecedent Address two open review comments on PR #1181: - Bump ios-e2e.yml xcuitest-gestures timeout-minutes 25 -> 30 to restore headroom after BE-0281's ~13min -> ~21min runtime growth on a gating job. - Fix 'runs it' -> 'runs both' in docs/architecture.md's idb entry: the bullet describes two things (back and device control), so the singular pronoun misread. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NgEYipfg6U2xLeCAxTTNF7

Wires idb and XCUITest actuation beyond the conformance job's `tap` into ios-e2e.yml: a new non-gating `actuation (idb)` job drives `back` (navigation.yaml) and device control — setLocation, clipboard, push (device.yaml + push.yaml) — on the Stable launch tab, the only surface idb can reach (idb collapses the native tab bar into one opaque group and cannot tap any tab, confirmed on-device before implementation). The existing `xcuitest (multi-touch)` job gains two more runs — search.yaml for `/type` and notices.yaml for `/swipe` + `/back` on the runner channel — reusing its already-built runner rather than a new job. Scope narrowed from the original proposal: idb cannot drive the tab-crossing gesture/text scenarios (gestures.yaml, controls.yaml) it assumed, so type/swipe/scroll/longPress/doubleTap stay idb-unproven on device pending a tab-free showcase screen (left for a follow-up item). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

…ntrol on-device - ios-e2e.yml: notices.yaml only uses `/swipe` + `/back` steps (no `/scroll` endpoint exists on the runner channel), so drop the stray "+ `/scroll`" from the job-header comment describing it. - docs/architecture.md and its docs/ja/architecture.md mirror: credit idb's `back` and device-control (setLocation/clipboard/push) actuation as now confirmed on-device via the `actuation (idb)` job (BE-0281), per CLAUDE.md's BE-0113 rule to keep architecture.md in step with behavior in the same PR. Addresses review feedback on PR #1181. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

…eout comment The Progress checklist bullet incorrectly lumped `tap` in with the actuators that stay on XCUITest, contradicting the PR's own actuation (idb) job (which runs navigation.yaml's `tap` step) and the Log entry below it, which already excludes `tap` correctly. Fix both the English and Japanese checklist bullets to match. Also update the xcuitest-gestures job's stale ~13min budget comment to ~21min, reflecting the two extra bajutsu run invocations (search.yaml, notices.yaml) this PR added to that job — timeout-minutes: 25 still has headroom. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

Address round-3 review feedback on PR #1181: the actuation (idb) job's three bajutsu-e2e steps (navigation/device/push) are independent scenarios bundled into one job for cost reasons, but a plain step failure skipped every step after it — hiding the other scenarios' pass/ fail signal and artifacts. Mark each step continue-on-error and add a final step that re-fails the job if any of the three failed, so all three always get a chance to run while the job-level result stays accurate. Also fixes documentation drift flagged by the same review round: the roadmap's References list still cited gestures.yaml/controls.yaml (never actually wired over idb) instead of the scenarios this PR wired (navigation/device/push/search/notices.yaml), and architecture.md overclaimed the actuation (idb) job "confirmed" back/device-control on-device before a passing run had been observed. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

…ture.md antecedent Address two open review comments on PR #1181: - Bump ios-e2e.yml xcuitest-gestures timeout-minutes 25 -> 30 to restore headroom after BE-0281's ~13min -> ~21min runtime growth on a gating job. - Fix 'runs it' -> 'runs both' in docs/architecture.md's idb entry: the bullet describes two things (back and device control), so the singular pronoun misread. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NgEYipfg6U2xLeCAxTTNF7

Addresses review feedback: consistent "never actuate" phrasing with the smoke job's description, a double-negative fix, a missing object, and a note on why the required xcuitest-gestures job skips the actuation job's continue-on-error pattern. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

0x0c had a problem deploying to claude-review July 17, 2026 13:49 — with GitHub Actions Error

github-actions Bot added the topic:dev-infra CI / build infrastructure (roadmap topic) label Jul 17, 2026

0x0c added a commit that referenced this pull request Jul 17, 2026

docs(roadmap): link BE-0281 to its implementing PR #1181

1fae2aa

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

0x0c temporarily deployed to claude-review July 17, 2026 13:51 — with GitHub Actions Inactive